Description

Implement the semantic state memory component from MemCoT (arXiv:2604.08216, #3564).

Add a SemanticStateAccumulator to TurnContext in zeph-core that:

• Summarizes the conceptual state after each LLM turn
• Accumulates semantic deltas across the conversation
• Feeds into ReasoningMemory retrieval as enriched query context

Config flag: memory.memcot.enabled (default: false). When disabled, accumulator is a no-op.

The accumulator must use a fast-tier provider (e.g., gpt-4o-mini) per the multi-model design rule — never a quality-tier provider. The distillation prompt must be short and cost-bounded.

Acceptance Criteria

• memory.memcot.enabled config flag wires/no-ops the accumulator
• Integration test on a multi-turn fixture shows accumulator content in debug dump
• No regression on single-turn benchmark
• Feature is completely invisible when enabled = false (no latency, no LLM calls)

References

• MemCoT paper: https://arxiv.org/abs/2604.08216
• Research synthesis: research(memory): MemCoT test-time memory scaling — multi-view long-context reasoning with episodic + semantic short-term memory #3564
• APEX-MEM spec: /specs/004-memory/004-7-memory-apex-magma.md §14.1
• Implementation notes: .local/testing/playbooks/memory-research-2026-04.md